src/backends/neon/NeonWorkloadFactory.cpp

   1 //
   2 // Copyright © 2017 Arm Ltd. All rights reserved.
   3 // SPDX-License-Identifier: MIT
   4 //
   5 #include "NeonWorkloadFactory.hpp"
   6 #include <armnn/Utils.hpp>
   7 #include <backends/CpuTensorHandle.hpp>
   8 #include <Layer.hpp>
   9
  10 #ifdef ARMCOMPUTENEON_ENABLED
  11 #include <arm_compute/runtime/Allocator.h>
  12
  13 #include <backends/MemCopyWorkload.hpp>
  14 #include "NeonTensorHandle.hpp"
  15 #include "workloads/NeonWorkloadUtils.hpp"
  16 #include "workloads/NeonWorkloads.hpp"
  17
  18 #include <backends/aclCommon/memory/IPoolManager.hpp>
  19 #endif
  20
  21 #include <backends/MakeWorkloadHelper.hpp>
  22
  23 #include <boost/polymorphic_cast.hpp>
  24
  25 namespace armnn
  26 {
  27
  28 bool NeonWorkloadFactory::IsLayerSupported(const Layer& layer, boost::optional<DataType> dataType,
  29                                            std::string& outReasonIfUnsupported)
  30 {
  31     return IWorkloadFactory::IsLayerSupported(Compute::CpuAcc, layer, dataType, outReasonIfUnsupported);
  32 }
  33
  34 #ifdef ARMCOMPUTENEON_ENABLED
  35
  36 NeonWorkloadFactory::NeonWorkloadFactory()
  37     : m_MemoryManager(std::make_unique<arm_compute::Allocator>(), BaseMemoryManager::MemoryAffinity::Offset)
  38 {
  39 }
  40
  41 std::unique_ptr<ITensorHandle> NeonWorkloadFactory::CreateSubTensorHandle(ITensorHandle& parent,
  42     TensorShape const& subTensorShape,
  43     unsigned int const* subTensorOrigin) const
  44 {
  45     BOOST_ASSERT(parent.GetType() == ITensorHandle::Neon);
  46
  47     const arm_compute::TensorShape shape = armcomputetensorutils::BuildArmComputeTensorShape(subTensorShape);
  48
  49     arm_compute::Coordinates coords;
  50     coords.set_num_dimensions(subTensorShape.GetNumDimensions());
  51     for (unsigned int i = 0; i < subTensorShape.GetNumDimensions(); i++)
  52     {
  53         // Arm compute indexes tensor coords in reverse order.
  54         unsigned int revertedIndex = subTensorShape.GetNumDimensions() - i - 1;
  55         coords.set(i, boost::numeric_cast<int>(subTensorOrigin[revertedIndex]));
  56     }
  57
  58     return std::make_unique<NeonSubTensorHandle>(
  59         boost::polymorphic_downcast<INeonTensorHandle*>(&parent), shape, coords);
  60 }
  61
  62 std::unique_ptr<ITensorHandle> NeonWorkloadFactory::CreateTensorHandle(const TensorInfo& tensorInfo) const
  63 {
  64     auto tensorHandle = std::make_unique<NeonTensorHandle>(tensorInfo);
  65     tensorHandle->SetMemoryGroup(m_MemoryManager.GetInterLayerMemoryGroup());
  66
  67     return tensorHandle;
  68 }
  69
  70 std::unique_ptr<ITensorHandle> NeonWorkloadFactory::CreateTensorHandle(const TensorInfo& tensorInfo,
  71                                                                        DataLayout dataLayout) const
  72 {
  73     auto tensorHandle = std::make_unique<NeonTensorHandle>(tensorInfo, dataLayout);
  74     tensorHandle->SetMemoryGroup(m_MemoryManager.GetInterLayerMemoryGroup());
  75
  76     return tensorHandle;
  77 }
  78
  79 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateInput(const InputQueueDescriptor& descriptor,
  80                                                             const WorkloadInfo&        info) const
  81 {
  82     return MakeWorkload<CopyMemGenericWorkload, CopyMemGenericWorkload>(descriptor, info);
  83 }
  84
  85 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateOutput(const OutputQueueDescriptor& descriptor,
  86                                                              const WorkloadInfo&        info) const
  87 {
  88     return MakeWorkload<CopyMemGenericWorkload, CopyMemGenericWorkload>(descriptor, info);
  89 }
  90
  91 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateActivation(const ActivationQueueDescriptor& descriptor,
  92                                                                  const WorkloadInfo&              info) const
  93 {
  94     return std::make_unique<NeonActivationWorkload>(descriptor, info);
  95 }
  96
  97 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateSoftmax(const SoftmaxQueueDescriptor& descriptor,
  98                                                               const WorkloadInfo&           info) const
  99 {
 100     return MakeWorkload<NeonSoftmaxFloatWorkload, NeonSoftmaxUint8Workload>(descriptor, info,
 101                                                                               m_MemoryManager.GetIntraLayerManager());
 102 }
 103
 104 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateSplitter(const SplitterQueueDescriptor& descriptor,
 105                                                                const WorkloadInfo&            info) const
 106 {
 107     return std::make_unique<NeonSplitterWorkload>(descriptor, info);
 108 }
 109
 110 std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateMerger(const MergerQueueDescriptor& descriptor,
 111                                                                     const WorkloadInfo&          info) const
 112 {
 113     return std::make_unique<NeonMergerWorkload>(descriptor, info);
 114 }
 115
 116 std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateFullyConnected(
 117     const FullyConnectedQueueDescriptor& descriptor, const WorkloadInfo& info) const
 118 {
 119     return MakeWorkload<NeonFullyConnectedWorkload, NeonFullyConnectedWorkload>(descriptor, info,
 120                                                                                 m_MemoryManager.GetIntraLayerManager());
 121 }
 122
 123 std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreatePermute(const PermuteQueueDescriptor& descriptor,
 124                                                                      const WorkloadInfo&           info) const
 125 {
 126     return std::make_unique<NeonPermuteWorkload>(descriptor, info);
 127 }
 128
 129 std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreatePooling2d(const Pooling2dQueueDescriptor& descriptor,
 130                                                                        const WorkloadInfo&           info) const
 131 {
 132     return std::make_unique<NeonPooling2dWorkload>(descriptor, info);
 133 }
 134
 135 std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateConvolution2d(
 136     const Convolution2dQueueDescriptor& descriptor, const WorkloadInfo& info) const
 137 {
 138     return std::make_unique<NeonConvolution2dWorkload>(descriptor, info,
 139                                                        m_MemoryManager.GetIntraLayerManager());
 140 }
 141
 142 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateDepthwiseConvolution2d(
 143     const DepthwiseConvolution2dQueueDescriptor& descriptor, const WorkloadInfo& info) const
 144 {
 145     return MakeWorkload<NeonDepthwiseConvolutionFloatWorkload, NeonDepthwiseConvolutionUint8Workload>(
 146         descriptor, info);
 147 }
 148
 149 std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateNormalization(
 150     const NormalizationQueueDescriptor& descriptor, const WorkloadInfo& info) const
 151 {
 152     return MakeWorkload<NeonNormalizationFloatWorkload, NullWorkload>(descriptor, info,
 153                                                                         m_MemoryManager.GetIntraLayerManager());
 154 }
 155
 156 std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateAddition(const AdditionQueueDescriptor& descriptor,
 157                                                                       const WorkloadInfo&            info) const
 158 {
 159     return MakeWorkload<NeonAdditionFloatWorkload, NullWorkload>(descriptor, info);
 160 }
 161
 162 std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateMultiplication(
 163     const MultiplicationQueueDescriptor& descriptor, const WorkloadInfo& info) const
 164 {
 165     return MakeWorkload<NeonMultiplicationFloatWorkload, NullWorkload>(descriptor, info);
 166 }
 167
 168 std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateDivision(
 169     const DivisionQueueDescriptor& descriptor, const WorkloadInfo& info) const
 170 {
 171     return MakeWorkload<NullWorkload, NullWorkload>(descriptor, info);
 172 }
 173
 174 std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateSubtraction(
 175     const SubtractionQueueDescriptor& descriptor, const WorkloadInfo& info) const
 176 {
 177     return MakeWorkload<NeonSubtractionFloatWorkload, NullWorkload>(descriptor, info);
 178 }
 179
 180 std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateBatchNormalization(
 181     const BatchNormalizationQueueDescriptor& descriptor, const WorkloadInfo& info) const
 182 {
 183     return MakeWorkload<NeonBatchNormalizationFloatWorkload, NullWorkload>(descriptor, info);
 184 }
 185
 186 std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateMemCopy(const MemCopyQueueDescriptor& descriptor,
 187                                                                      const WorkloadInfo&        info) const
 188 {
 189     if (descriptor.m_Inputs.empty() || !descriptor.m_Inputs[0])
 190     {
 191         throw InvalidArgumentException("NeonWorkloadFactory: Invalid null input for MemCopy workload");
 192     }
 193
 194     return MakeWorkload<CopyMemGenericWorkload, CopyMemGenericWorkload>(descriptor, info);
 195 }
 196
 197 std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateResizeBilinear(
 198     const ResizeBilinearQueueDescriptor& descriptor,
 199     const WorkloadInfo& info) const
 200 {
 201     return nullptr;
 202 }
 203
 204 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateFakeQuantization(
 205     const FakeQuantizationQueueDescriptor& descriptor,
 206     const WorkloadInfo& info) const
 207 {
 208     return nullptr;
 209 }
 210
 211 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateL2Normalization(const L2NormalizationQueueDescriptor& descriptor,
 212     const WorkloadInfo& info) const
 213 {
 214     return MakeWorkload<NeonL2NormalizationFloatWorkload, NullWorkload>(descriptor, info,
 215                                                                           m_MemoryManager.GetIntraLayerManager());
 216 }
 217
 218 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateConstant(const ConstantQueueDescriptor& descriptor,
 219     const WorkloadInfo& info) const
 220 {
 221     return std::make_unique<NeonConstantWorkload>(descriptor, info);
 222 }
 223
 224 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateReshape(const ReshapeQueueDescriptor& descriptor,
 225     const WorkloadInfo& info) const
 226 {
 227     return std::make_unique<NeonReshapeWorkload>(descriptor, info);
 228 }
 229
 230 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateFloor(const FloorQueueDescriptor& descriptor,
 231     const WorkloadInfo& info) const
 232 {
 233     return MakeWorkload<NeonFloorFloatWorkload, NullWorkload>(descriptor, info);
 234 }
 235
 236 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateLstm(const LstmQueueDescriptor& descriptor,
 237     const WorkloadInfo& info) const
 238 {
 239     return MakeWorkload<NeonLstmFloatWorkload, NullWorkload>(descriptor, info);
 240 }
 241
 242 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateConvertFp16ToFp32(
 243     const ConvertFp16ToFp32QueueDescriptor& descriptor,
 244     const WorkloadInfo& info) const
 245 {
 246     return std::make_unique<NeonConvertFp16ToFp32Workload>(descriptor, info);
 247 }
 248
 249 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateConvertFp32ToFp16(
 250     const ConvertFp32ToFp16QueueDescriptor& descriptor,
 251     const WorkloadInfo& info) const
 252 {
 253     return std::make_unique<NeonConvertFp32ToFp16Workload>(descriptor, info);
 254 }
 255
 256 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateMean(const MeanQueueDescriptor& descriptor,
 257                                                            const WorkloadInfo& info) const
 258 {
 259     return MakeWorkload<NullWorkload, NullWorkload>(descriptor, info);
 260 }
 261
 262 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreatePad(const PadQueueDescriptor& descriptor,
 263                                                           const WorkloadInfo& info) const
 264 {
 265     return MakeWorkload<NullWorkload, NullWorkload>(descriptor, info);
 266 }
 267
 268 void NeonWorkloadFactory::Finalize()
 269 {
 270     m_MemoryManager.Finalize();
 271 }
 272
 273 void NeonWorkloadFactory::Release()
 274 {
 275     m_MemoryManager.Release();
 276 }
 277
 278 void NeonWorkloadFactory::Acquire()
 279 {
 280     m_MemoryManager.Acquire();
 281 }
 282
 283 #else // Compiled without ArmCompute libs
 284
 285 NeonWorkloadFactory::NeonWorkloadFactory()
 286 {
 287 }
 288
 289 std::unique_ptr<ITensorHandle> NeonWorkloadFactory::CreateSubTensorHandle(ITensorHandle& parent,
 290     TensorShape const& subTensorShape,
 291     unsigned int const* subTensorOrigin) const
 292 {
 293     return nullptr;
 294 }
 295
 296 std::unique_ptr<ITensorHandle> NeonWorkloadFactory::CreateTensorHandle(const TensorInfo& tensorInfo) const
 297 {
 298     return nullptr;
 299 }
 300
 301 std::unique_ptr<ITensorHandle> NeonWorkloadFactory::CreateTensorHandle(const TensorInfo& tensorInfo,
 302                                                                        DataLayout dataLayout) const
 303 {
 304     return nullptr;
 305 }
 306
 307 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateInput(const InputQueueDescriptor& descriptor,
 308                                                             const WorkloadInfo&        info) const
 309 {
 310     return nullptr;
 311 }
 312
 313 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateOutput(const OutputQueueDescriptor& descriptor,
 314                                                              const WorkloadInfo&        info) const
 315 {
 316     return nullptr;
 317 }
 318
 319 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateActivation(const ActivationQueueDescriptor& descriptor,
 320                                                                  const WorkloadInfo&              info) const
 321 {
 322     return nullptr;
 323 }
 324
 325 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateSoftmax(const SoftmaxQueueDescriptor& descriptor,
 326                                                               const WorkloadInfo&           info) const
 327 {
 328     return nullptr;
 329 }
 330
 331 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateSplitter(const SplitterQueueDescriptor& descriptor,
 332                                                                const WorkloadInfo&            info) const
 333 {
 334     return nullptr;
 335 }
 336
 337 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateMerger(const MergerQueueDescriptor& descriptor,
 338                                                              const WorkloadInfo&          info) const
 339 {
 340     return nullptr;
 341 }
 342
 343 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateFullyConnected(const FullyConnectedQueueDescriptor& descriptor,
 344                                                                      const WorkloadInfo&                  info) const
 345 {
 346     return nullptr;
 347 }
 348
 349 std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreatePermute(const PermuteQueueDescriptor& descriptor,
 350                                                                      const WorkloadInfo&           info) const
 351 {
 352     return nullptr;
 353 }
 354
 355 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreatePooling2d(const Pooling2dQueueDescriptor& descriptor,
 356                                                                 const WorkloadInfo&           info) const
 357 {
 358     return nullptr;
 359 }
 360
 361 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateConvolution2d(const Convolution2dQueueDescriptor& descriptor,
 362                                                                     const WorkloadInfo&               info) const
 363 {
 364     return nullptr;
 365 }
 366
 367 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateDepthwiseConvolution2d(
 368     const DepthwiseConvolution2dQueueDescriptor& descriptor, const WorkloadInfo& info) const
 369 {
 370     return nullptr;
 371 }
 372
 373 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateNormalization(const NormalizationQueueDescriptor& descriptor,
 374                                                                     const WorkloadInfo&                 info) const
 375 {
 376     return nullptr;
 377 }
 378
 379 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateAddition(const AdditionQueueDescriptor& descriptor,
 380                                                                const WorkloadInfo&            info) const
 381 {
 382     return nullptr;
 383 }
 384
 385 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateBatchNormalization(const BatchNormalizationQueueDescriptor& data,
 386                                                                          const WorkloadInfo& info) const
 387 {
 388     return nullptr;
 389 }
 390
 391 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateMultiplication(const MultiplicationQueueDescriptor& data,
 392                                                                      const WorkloadInfo&                  info) const
 393 {
 394     return nullptr;
 395 }
 396
 397 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateMemCopy(const MemCopyQueueDescriptor& descriptor,
 398                                                               const WorkloadInfo&        info) const
 399 {
 400     return nullptr;
 401 }
 402
 403 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateResizeBilinear(const ResizeBilinearQueueDescriptor& descriptor,
 404                                                                      const WorkloadInfo& info) const
 405 {
 406     return nullptr;
 407 }
 408
 409 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateFakeQuantization(
 410         const FakeQuantizationQueueDescriptor& descriptor, const WorkloadInfo& info) const
 411 {
 412     return nullptr;
 413 }
 414
 415 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateL2Normalization(const L2NormalizationQueueDescriptor& descriptor,
 416     const WorkloadInfo& info) const
 417 {
 418     return nullptr;
 419 }
 420
 421 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateConstant(const ConstantQueueDescriptor& descriptor,
 422     const WorkloadInfo& info) const
 423 {
 424     return nullptr;
 425 }
 426
 427 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateReshape(const ReshapeQueueDescriptor& descriptor,
 428     const WorkloadInfo&           info) const
 429 {
 430     return nullptr;
 431 }
 432
 433 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateFloor(const FloorQueueDescriptor& descriptor,
 434     const WorkloadInfo& info) const
 435 {
 436     return nullptr;
 437 }
 438
 439 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateLstm(const LstmQueueDescriptor& descriptor,
 440     const WorkloadInfo& info) const
 441 {
 442     return nullptr;
 443 }
 444
 445 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateConvertFp16ToFp32(
 446     const ConvertFp16ToFp32QueueDescriptor& descriptor,
 447     const WorkloadInfo& info) const
 448 {
 449     return nullptr;
 450 }
 451
 452 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateConvertFp32ToFp16(
 453     const ConvertFp32ToFp16QueueDescriptor& descriptor,
 454     const WorkloadInfo& info) const
 455 {
 456     return nullptr;
 457 }
 458
 459 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateDivision(const DivisionQueueDescriptor& data,
 460                                                                const WorkloadInfo& info) const
 461 {
 462     return nullptr;
 463 }
 464
 465 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateSubtraction(const SubtractionQueueDescriptor& data,
 466                                                                   const WorkloadInfo& info) const
 467 {
 468     return nullptr;
 469 }
 470
 471 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateMean(const MeanQueueDescriptor& descriptor,
 472                                                            const WorkloadInfo& info) const
 473 {
 474     return nullptr;
 475 }
 476
 477 std::unique_ptr<IWorkload> NeonWorkloadFactory::CreatePad(const PadQueueDescriptor& descriptor,
 478                                                           const WorkloadInfo& info) const
 479 {
 480     return nullptr;
 481 }
 482
 483 void NeonWorkloadFactory::Finalize()
 484 {}
 485
 486 void NeonWorkloadFactory::Release()
 487 {}
 488
 489 void NeonWorkloadFactory::Acquire()
 490 {}
 491
 492 #endif
 493
 494 } //namespace armnn